Skip to content
Back to Project Archive

Infinite Intelligence

BYOK multi-agent council v4.0.0 — parallel/sequential/round-robin, HITL, branching, PDF export via gemma-4-31b-it. Client-only, no backend.

Infinite Intelligence Logo
// PROJECT_ID
the-infinite-intelligence
// AUTHOR
DHAATRIK CHOWDHURY
// SUBJECT
CASE STUDY
// CLASSIFICATION
PUBLIC // OPEN
// TELEMETRY
STATUS: OPERATIONAL // MODEL: GEMMA_4_31B // BACKEND: NONE
// IMPACT.LOG
// FUCKUP_LOG
RECORDED
[ FUCKUP LOG ]  I ran an unbounded multi-agent debate test without depth caps and woke up to a rate-limited API key — v4 adds token/cost visibility but you still own the guardrails.

SYS.STATUS: v4.0.0 operational — council orchestration, HITL steering, executive export

The Infinite Intelligence is a client-side multi-agent orchestration SPA. You supply a Gemini API key; the app deploys a four-agent default squad (plus Beta Mode dynamic personas), runs parallel, sequential, or round-robin flows with critique rounds, optional human-in-the-loop steering, session branching, artifact extraction, and PDF/Markdown export.

Model per README: gemma-4-31b-it via @google/genai. No backend — prompts and keys stay in your browser session.

This page updates portfolio copy that undersold v4 as “three agents” and tagged Tailwind without a matching package.json dep.

What it is (scope)

SurfaceWhat you do there
ModesParallel, sequential, round-robin (up to 5 critique rounds)
First principlesPre-phase deconstructs prompt into constraints
Default squadFour configured expert personas
Beta ModeTask-tailored dynamic agent generation
HITLThumbs up/down, steering comments mid-run
BranchingFork history at any node
ArtifactsSide panel for code/JSON snippets
ExportPDF (html2canvas + jsPDF) + Markdown reports
MetricsToken and estimated cost display

Stack: React 19.2, TypeScript 5.8, Vite 6, @google/genai, react-markdown, remark-gfm, Motion, html2canvas, jsPDF, Vitest.

Requires VITE_GEMINI_API_KEY. MIT licensed.

Who I built it for

  • Skeptics of single-shot LLM certainty who want structured critique
  • Developers studying client-side agent orchestration patterns
  • Teams prototyping council-of-models UX before building server infra

Not for: free inference (you pay API usage), unattended infinite debate, or enterprise persistence without adding a backend.

Fuckups & learnings

  • Undersold v4 on portfolio. Four agents + Beta + HITL + branching beats “three debaters” fiction.
  • Unbounded loops cost real money. Rate-limit bill fuckup — cap rounds, use HITL, watch metrics.
  • Debate amplifies noise when Critic prompts are vague.
  • README vs package.json drift. Tailwind mentioned in README; v4.0.0 deps are React/Vite/genai/export stack — verify before tagging.

Honest limitations

LimitationReality
BYOK costsToken usage scales with agents × rounds × prompt size
No server persistenceExport reports; refresh clears in-tab session
Model id driftKeep README and constants.ts aligned
LatencyCouncil modes are slower than single-shot chat

Deep-dive transmissions

Read in order:

  1. Why a council of models
  2. Client orchestration tech stack
  3. BYOK economics and guardrails

Run it locally

git clone https://github.com/dhaatrik/the-infinite-intelligence.git
cd the-infinite-intelligence
npm install
# .env: VITE_GEMINI_API_KEY=your_key
npm run dev

Tests: npm run test — Vitest.

Closing transmission

Process over bravado. Clone it, cap your rounds, break a workflow, see if the council catches what one shot missed. Start with why-and-what.